Drone detection method based on machine learning

Abstract

The present invention discloses a drone detection method based on machine learning, comprising: data preprocessing: performing data filtering processing on drone frequency band spectrum data collected by a receiver; signal extraction: performing a signal extraction operation on the preprocessed data to extract all signals with similar bandwidth to the drone signal; data normalization: performing z-score normalization processing on all extracted signals; feature extraction: performing a feature extraction operation on the normalized signal data to serve as different data features for distinguishing drone signals from non-drone signals; drone signal detection: using a machine learning algorithm to detect drone signals based on each characteristic value of the given signal, and outputting drone signal parameters if a drone signal is found. The present invention combines the machine learning algorithm to achieve drone signal detection and parameter measurement.

Description

Technical Field

[0001] The present invention relates to the field of signal processing technology, and in particular to a drone detection method based on machine learning. Background Art

[0002] There are currently three main methods for detecting drones: radar-based drone detection, audio and video-based drone detection, and radio-based drone detection. The advantages and disadvantages of each detection method are shown in Table 1.

[0003] Table 1 Advantages and disadvantages of various detection methods

[0004]

[0005] Using radar to detect drone signals is currently the most comprehensive solution, but radar detection systems are too expensive and are only effective against large, high-speed moving targets. Drones are low-speed, low-altitude flying platforms with a small radar cross-section, known as low-speed, small-size (LSS) platforms. Radar detection methods have insufficient detection efficiency for drone targets. Another disadvantage of radar is its high power consumption and inability to operate around the clock. Due to certain limitations of audio and video detection equipment, audio and video detection methods have a short range. The environment also has a significant impact on audio and video detection. For example, cameras cannot capture drone images in foggy weather, and microphones cannot capture drone audio in noisy environments. Radio detection methods detect drone signals through spectrum detection and use radio frequency sensors to collect drone communication signals. Its advantages are low cost, low computational complexity, and no echo interference. Summary of the Invention

[0006] In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a drone detection method based on machine learning. The present invention combines machine learning algorithms to realize the detection and parameter measurement of drone signals.

[0007] To achieve the above objectives, the present invention adopts a technical solution: a drone detection method based on machine learning, comprising the following steps:

[0008] Step 1: Data preprocessing: Perform data filtering on the UAV frequency band spectrum data collected by the receiver;

[0009] Step 2: Signal extraction: Perform signal extraction on the preprocessed data to extract all signals with similar bandwidth to the drone signal.

[0010] Step 3, data standardization: perform z-score standardization on all extracted signals;

[0011] Step 4: Feature extraction: Perform feature extraction on the normalized signal data to serve as the different data features to distinguish UAV signals from non-UAV signals.

[0012] Step 5: Drone signal detection: Based on the given signal eigenvalues, a machine learning algorithm is used to detect drone signals. If a drone signal is found, the drone signal parameters are output.

[0013] As a further improvement of the present invention, in step 1, median filtering is used to perform data filtering processing on the drone frequency band spectrum data collected by the receiver, specifically as follows:

[0014] Assume that the collected UAV frequency band spectrum data is x(iN),…,x(i),x(i+N), define a window of odd length L, L=2N+1, N is a positive integer, sort the L signal samples, and take the sample value at i as the output value of the median filter: y(i)=Med[x(iN),…,x(i),x(i+n)].

[0015] As a further improvement of the present invention, the step 2 is specifically as follows:

[0016] Define the frequency step (Khz) of the frequency band data, find all signals with a bandwidth greater than 5Mhz from the entire frequency band, and define the signal set extracted from the frequency band Where n is the number of signals with a bandwidth greater than 5 MHz in the frequency band, and a single signal is represented by Where k is the total number of sampling points of a single signal.

[0017] As a further improvement of the present invention, in step 3, the standardized formula is: Get the original sequence The average value μ, the standard deviation σ of the original sequence, and the standardized sequence are obtained

[0018] As a further improvement of the present invention, in step 4, the extracted features include signal bandwidth, coefficient of variation, kurtosis, and volatility.

[0019] As a further improvement of the present invention, the calculation formulas for the coefficient of variation, kurtosis, and volatility are as follows:

[0020] Define μ as a sequence The average value of the sequence The standard deviation of n is the total number of sample points in the sequence:

[0021] Define the coefficient of variation as CV:

[0022]

[0023] Define kurtosis as K:

[0024]

[0025] Define volatility as wave rate:

[0026]

[0027] Y = quantile(X, p) is the percentile calculation, where X is a vector or matrix, and p is the percentile in the range [0, 1]. The mathematical meaning is: for vector X, P{X <= Y} = p.

[0028] As a further improvement of the present invention, the step 5 is specifically as follows:

[0029] The extracted drone image transmission signals and non-drone feature data are integrated together, and the drone signals and non-drone signals are labeled to form a data set. The data set is put into the random forest algorithm in machine learning for detection to obtain the detection results.

[0030] The beneficial effects of the present invention are:

[0031] This invention combines the characteristics of drone signals with machine learning algorithms to greatly reduce the false alarm rate of drone signal detection and solve the problem of detecting illegal drone signals in complex environments. Experiments have shown that this algorithm can detect drone signals in interference environments and has extremely high recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flowchart of an embodiment of the present invention;

[0033] Figure 2 This is a simulation diagram of spectrum data before filtering in an embodiment of the present invention;

[0034] Figure 3 This is a simulation diagram after filtering of spectrum data in an embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram of drone signals in an embodiment of the present invention;

[0036] Figure 5 Schematic diagram of non-UAV signals in an embodiment of the present invention;

[0037] Figure 6 Schematic diagram of bandwidth characteristics in an embodiment of the present invention;

[0038] Figure 7 A schematic diagram of defining variation features in an embodiment of the present invention;

[0039] Figure 8Schematic diagram of kurtosis characteristics in an embodiment of the present invention;

[0040] Figure 9 Schematic diagram of volatility characteristics in an embodiment of the present invention;

[0041] Figure 10 This is a distribution diagram of the number of trees and the out-of-bag error rate in the random forest algorithm in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0043] Example

[0044] like Figure 1 As shown in the figure, a drone detection method based on machine learning mainly refers to an algorithm for detecting drone image transmission signals within a specified range, including:

[0045] S1. Data Preprocessing: Filter the 2.4-2.5 GHz drone frequency band spectrum data collected by the receiver to remove noise. Common smoothing filtering methods include low-pass filtering, median filtering, and mean filtering. After comparison, median filtering was found to be the most effective smoothing method, so it was selected as the signal smoothing method.

[0046] Specifically, define the original spectrum sequence as x(iN),…,x(i),x(i+N). Define a window of odd length L, (L=2N+1, N is a positive integer). Sort the L signal samples and take the sample value at position i as the output value of the median filter, such as y(i)=Med[x(iN),…,x(i),x(i+n)]. The simulation diagram before and after filtering is as follows Figure 2 and Figure 3 As shown:

[0047] from Figure 2 and Figure 3 It can be seen that there are too many signal glitches before filtering, which interferes with subsequent signal extraction. After median filtering, the signal curve becomes smoother.

[0048] S2. Signal extraction: Perform signal extraction on the preprocessed data to extract all signals with similar bandwidth to the drone signal.

[0049] Specifically, define the frequency step (Khz) of the frequency band data, find all signals with a bandwidth greater than 5Mhz from the entire frequency band, and define the signal set extracted from the frequency band Where n is the number of signals with a bandwidth greater than 5 MHz in the frequency band, and a single signal is represented by Among them, k is the total number of sampling points of a single signal. The extracted UAV signal and non-UAV signal simulation diagram are as follows Figure 4 and Figure 5 shown.

[0050] from Figure 4 and Figure 5 It can be seen that the spectrum curve of the drone signal is steeper on both sides of the spectrum than the non-drone signal envelope, and is flat within the band.

[0051] S3. Data standardization: In order to unify the comparison standards and ensure the reliability of the results, all extracted signals are z-score standardized.

[0052] Specifically, to normalize the input signal data, you need to use its mean and standard deviation, mapping the standard deviation of the original signal data to 1. After processing, the data conforms to the normal distribution (mean is 0, standard deviation is 1). This is also called standard deviation normalization. The normalization formula is as follows:

[0053] Get the original sequence The average value μ, the standard deviation σ of the original sequence, and the standardized sequence are obtained

[0054] S4. Feature Extraction: By extracting a sufficient number of features from the data, we can minimize the problem of overfitting in machine learning classifiers. We perform feature extraction on the standardized signal data, including signal bandwidth, coefficient of variation, kurtosis, and volatility, as distinct data features to distinguish UAV signals from non-UAV signals.

[0055] Specifically, bandwidth is a basic characteristic of a signal; the coefficient of variation is used to measure the degree of data divergence and is the ratio of the data standard deviation to the mean; kurtosis is used to measure whether the data distribution is more peaked or flatter than the normal distribution. High kurtosis data has a clear peak near the mean, drops rapidly and has a heavy tail, while low kurtosis data often has a flat top near the mean; volatility is defined as the 90th percentile - 10th percentile after a certain data standardization;

[0056] The calculation formulas for the coefficient of variation, kurtosis, and volatility are as follows:

[0057] Define μ as a sequence The average value of the sequence The standard deviation of n is the total number of sample points in the sequence:

[0058] Define the coefficient of variation as CV:

[0059]

[0060] Define kurtosis as K:

[0061]

[0062] Define volatility as wave rate:

[0063]

[0064] Y = quantile(X, p) is the percentile calculation, where X is a vector or matrix, and p is the percentile in the range [0, 1]. The mathematical meaning is: for vector X, P{X <= Y} = p.

[0065] Specific features Figure 6-Figure 9 As shown, Figure 6-Figure 9 Four characteristic statistical curves of drone signals and non-drone signals are given. It can be seen from the figure that detecting drone signals based on a single feature will result in false alarms, so it is necessary to combine multi-dimensional features for discrimination.

[0066] S5. Drone signal detection: Based on the given signal eigenvalues, a machine learning algorithm is used to detect drone signals. If a drone signal is found, the drone signal parameters are output.

[0067] Specifically, the extracted feature data of drone image transmission signals and non-drone signals are integrated, and the drone and non-drone signals are labeled to form a dataset. This dataset is then tested using the random forest algorithm in machine learning to obtain the test results. The distribution of the number of trees and the out-of-bag error rate in the random forest algorithm is shown in Figure 10.

[0068] In machine learning, there can still be some discrepancy between predictions and actual results. Therefore, this example uses accuracy, precision, and recall to evaluate the quality of a classifier. Classification is divided into two categories: positive and negative. TP indicates a drone is predicted to be a drone, TN indicates a non-drone is predicted to be a non-drone, FP indicates a non-drone is predicted to be a drone, and FN indicates a drone is predicted to be a non-drone.

[0069] Accuracy: This is the percentage of all correct predictions (the sum of drones predicted as drones and non-drones predicted as non-drones) to all predictions. The formula is defined as follows:

[0070]

[0071] Precision: Also known as the precision rate, this is the ratio of drones predicted to all drones (the sum of drones predicted to be drones and non-drones predicted to be drones). This is the ratio of actual drone predictions to all drone predictions. The formula is defined as follows:

[0072]

[0073] Recall: Also known as recall, it is the ratio of drones predicted to actual drones (the sum of drones predicted to be drones and drones predicted to be non-drones). It is the ratio of actual drones predicted to actual drones. The formula is defined as follows:

[0074]

[0075] Since the amount of non-UAV data in the dataset is much larger than that of UAV data, the harmonic mean F is used. score , the formula definition is as follows:

[0076]

[0077] Using the statistical indicator probability of detection (POD), the formula definition is as follows:

[0078]

[0079] Cumulatives success index (CSI), the formula is defined as follows:

[0080]

[0081]

[0082] The above-described embodiments merely represent specific implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. A drone detection method based on machine learning, characterized in that: The following steps are involved: Step 1: Data preprocessing: Perform data filtering on the UAV frequency band spectrum data collected by the receiver; Step 2: Signal extraction: Perform signal extraction on the preprocessed data to extract all signals with similar bandwidth to the drone signal. Step 3, data standardization: perform z-score standardization on all extracted signals; Step 4: Feature extraction: Perform feature extraction on the normalized signal data to serve as the different data features to distinguish UAV signals from non-UAV signals. In step 4, the extracted features include signal bandwidth, coefficient of variation, kurtosis, and volatility; The calculation formulas for the coefficient of variation, kurtosis, and volatility are as follows: Define μ as a sequence The average value of the sequence The standard deviation of n is the total number of sample points in the sequence: Define the coefficient of variation as CV: Define kurtosis as K: Define volatility as wave rate: Among them, Y = quantile(X, p) is the percentile calculation, where X is a vector or matrix, p is the percentile, and the value range is [0, 1]. The mathematical meaning is: for vector X, P{X <= Y} = p; Step 5: Drone signal detection: Based on the given signal eigenvalues, a machine learning algorithm is used to detect drone signals. If a drone signal is found, the drone signal parameters are output.

2. The drone detection method based on machine learning according to claim 1, characterized in that: In step 1, median filtering is used to perform data filtering on the drone frequency band spectrum data collected by the receiver, as follows: Assume that the collected UAV frequency band spectrum data is x(iN),…,x(i),x(i+N), define a window of odd length L, L=2N+1, N is a positive integer, sort the L signal samples, and take the sample value at i as the output value of the median filter: y(i)=Med[x(iN),…,x(i),x(i+n)].

3. The drone detection method based on machine learning according to claim 1 or 2, characterized in that: The step 2 is specifically as follows: Define the frequency step of the frequency band data, find all signals with a bandwidth greater than 5Mhz from the entire frequency band, and define the signal set extracted from the frequency band Where n is the number of signals with a bandwidth greater than 5 MHz in the frequency band, and a single signal is represented by Where k is the total number of sampling points of a single signal.

4. The drone detection method based on machine learning according to claim 3, characterized in that: In step 3, the normalization formula is: Get the original sequence The average value μ, the standard deviation σ of the original sequence, and the standardized sequence are obtained 5. The drone detection method based on machine learning according to claim 1, characterized in that: The step 5 is specifically as follows: The extracted drone image transmission signals and non-drone feature data are integrated together, and the drone signals and non-drone signals are labeled to form a data set. The data set is put into the random forest algorithm in machine learning for detection to obtain the detection results.

Images